summaryrefslogtreecommitdiffstats
path: root/modules/watchdog
diff options
context:
space:
mode:
Diffstat (limited to 'modules/watchdog')
-rw-r--r--modules/watchdog/.packaging/test.config4
-rw-r--r--modules/watchdog/README.rst43
-rw-r--r--modules/watchdog/watchdog.lua137
3 files changed, 184 insertions, 0 deletions
diff --git a/modules/watchdog/.packaging/test.config b/modules/watchdog/.packaging/test.config
new file mode 100644
index 0000000..9d1a291
--- /dev/null
+++ b/modules/watchdog/.packaging/test.config
@@ -0,0 +1,4 @@
+-- SPDX-License-Identifier: GPL-3.0-or-later
+modules.load('watchdog')
+assert(watchdog)
+quit()
diff --git a/modules/watchdog/README.rst b/modules/watchdog/README.rst
new file mode 100644
index 0000000..514f6c0
--- /dev/null
+++ b/modules/watchdog/README.rst
@@ -0,0 +1,43 @@
+.. SPDX-License-Identifier: GPL-3.0-or-later
+
+.. _mod-watchdog:
+
+Watchdog
+========
+
+This module cooperates with Systemd watchdog to restart the process in case
+the internal event loop gets stuck. The upstream Systemd unit files are configured
+to use this feature, which is turned on with the ``WatchdogSec=`` directive
+in the service file.
+
+As an optional feature, this module can also do an internal DNS query to check if resolver
+answers correctly. To use this feature you must configure DNS name and type to query for:
+
+.. code-block:: lua
+
+ watchdog.config({ qname = 'nic.cz.', qtype = kres.type.A })
+
+Each single query from watchdog must result in answer with
+RCODE = NOERROR or NXDOMAIN. Any other result will terminate the resolver
+(with SIGABRT) to allow the supervisor process to do cleanup, gather coredump
+and restart the resolver.
+
+It is recommended to use a name with a very short TTL to make sure the watchdog
+is testing all parts of resolver and not only its cache. Obviously this check
+makes sense only when used with very reliable domains; otherwise a failure
+on authoritative side will shutdown resolver!
+
+`WatchdogSec` specifies deadline for supervisor when the process will be killed.
+Watchdog queries are executed each `WatchdogSec / 2` seconds.
+This implies that **half** of `WatchdogSec` interval must be long enough for
+normal DNS query to succeed, so do not forget to add two or three seconds
+for random network timeouts etc.
+
+The module is loaded by default. If you'd like to disable it you can unload it:
+
+.. code-block:: lua
+
+ modules.unload('watchdog')
+
+Beware that unloading the module without disabling watchdog feature in supervisor
+will lead to infinite restart loop.
diff --git a/modules/watchdog/watchdog.lua b/modules/watchdog/watchdog.lua
new file mode 100644
index 0000000..710b8a8
--- /dev/null
+++ b/modules/watchdog/watchdog.lua
@@ -0,0 +1,137 @@
+-- SPDX-License-Identifier: GPL-3.0-or-later
+local ffi = require('ffi')
+
+ffi.cdef([[
+ int sd_watchdog_enabled(int unset_environment, uint64_t *usec);
+ int sd_notify(int unset_environment, const char *state);
+ void abort(void);
+]])
+
+local watchdog = {}
+local private = {}
+
+local function sd_signal_ok()
+ ffi.C.sd_notify(0, 'WATCHDOG=1')
+end
+
+function private.fail_callback()
+ log('[watchdog] ABORTING resolver, supervisor is expected to restart it')
+ ffi.C.abort()
+end
+
+-- logging
+local function add_tracer(logbuf)
+ return function (req)
+ local function qrylogger(_, msg)
+ jit.off(true, true) -- JIT for (C -> lua)^2 nesting isn't allowed
+ table.insert(logbuf, ffi.string(msg))
+ end
+ req.trace_log = ffi.cast('trace_log_f', qrylogger)
+ end
+end
+
+local function check_answer(logbuf)
+ return function (pkt, req)
+ req.trace_log:free()
+ if pkt ~= nil and (pkt:rcode() == kres.rcode.NOERROR
+ or pkt:rcode() == kres.rcode.NXDOMAIN) then
+ private.ok_callback()
+ return
+ end
+ log('[watchdog] watchdog query returned unexpected answer! query verbose log:')
+ log(table.concat(logbuf, ''))
+ if pkt ~= nil then
+ log('[watchdog] problematic answer:\n%s', pkt)
+ else
+ log('[watchdog] answer was dropped')
+ end
+ -- failure! quit immediatelly to allow process supervisor to restart us
+ private.fail_callback()
+ end
+end
+private.check_answer_callback = check_answer
+
+local function timer()
+ local logbuf = {}
+ -- fire watchdog query
+ if private.qname and private.qtype then
+ if verbose() then
+ log('[watchdog] starting watchdog query %s %s', private.qname, private.qtype)
+ end
+ resolve(private.qname,
+ private.qtype,
+ kres.class.IN,
+ {'TRACE'},
+ private.check_answer_callback(logbuf),
+ add_tracer(logbuf))
+ else
+ private.ok_callback()
+ end
+end
+
+function watchdog.config(cfg)
+ -- read only
+ if not cfg then
+ return private
+ end
+
+ local interval = tonumber(cfg.interval or private.interval or 10000)
+ if not interval or interval < 1 then
+ error('[watchdog] interval must be >= 1 ms')
+ end
+ private.interval = interval
+
+ -- qname = nil will disable DNS queries
+ private.qname = cfg.qname
+ private.qtype = cfg.qtype or kres.type.A
+
+ -- restart timers
+ watchdog.deinit()
+ private.event = event.recurrent(private.interval, timer)
+ return private
+end
+
+-- automatically enable watchdog if it is configured in systemd
+function watchdog.init()
+ if private.event then
+ error('[watchdog] module is already loaded')
+ end
+ local timeoutptr = ffi.new('uint64_t[1]')
+ local systemd_present, ret = pcall(function() return ffi.C.sd_watchdog_enabled(0, timeoutptr) end)
+ if not systemd_present then
+ if verbose() then
+ log('[watchdog] systemd library not detected')
+ end
+ return
+ end
+ private.ok_callback = sd_signal_ok
+ if ret < 0 then
+ error('[watchdog] %s', ffi.string(ffi.C.knot_strerror(math.abs(ret))))
+ return
+ elseif ret == 0 then
+ if verbose() then
+ log('[watchdog] disabled in systemd (WatchdogSec= not specified)')
+ end
+ return
+ end
+ local timeout = tonumber(timeoutptr[0]) / 1000 -- convert to ms
+ local interval = timeout / 2 -- halve interval to make sure we are never late
+ if interval < 1 then
+ log('[watchdog] error: WatchdogSec= must be at least 2ms! (got %d usec)',
+ tonumber(timeoutptr[0]))
+ end
+ watchdog.config({ interval = interval })
+ if verbose() then
+ log('[watchdog] systemd watchdog enabled (check interval: %s ms, timeout: %s ms)',
+ private.interval, timeout)
+ end
+end
+
+function watchdog.deinit()
+ if private.event then
+ event.cancel(private.event)
+ private.event = nil
+ end
+end
+
+return watchdog